fix(security): harden sandbox SSH with mandatory HMAC secret, NetworkPolicy, and nonce replay detection (#127)

* fix(security): harden sandbox SSH with mandatory HMAC secret, NetworkPolicy, and nonce replay detection

Closes #25

- Make NEMOCLAW_SSH_HANDSHAKE_SECRET mandatory: server and sandbox both
  refuse to start if the secret is empty/unset. Cluster deployments
  auto-generate it via openssl rand in the entrypoint script.
- Add Kubernetes NetworkPolicy restricting sandbox port 2222 ingress to
  the gateway pod only, preventing lateral movement from other cluster
  workloads.
- Add NSSH1 nonce replay detection with a TTL-bounded cache, rejecting
  replayed handshakes within the timestamp validity window.
- Add unit tests for verify_preface (valid, replay, expired, bad HMAC,
  malformed) and env injection.

* fix(deploy): pass sshHandshakeSecret in fast deploy helm upgrade

---------

Co-authored-by: John Myers <johntmyers@users.noreply.github.com>

architecture/sandbox-connect.md

@@ -345,7 +345,13 @@ NSSH1 <token> <timestamp> <nonce> <hmac>\n
 3. Reject if skew exceeds `ssh_handshake_skew_secs` (default: 300 seconds)
 4. Recompute HMAC-SHA256 over `token|timestamp|nonce` with the shared secret
 5. Compare computed signature against the received signature (constant-time via `hmac` crate)
-6. Respond with `OK\n` on success or `ERR\n` on failure
+6. Check nonce against the replay cache; reject if the nonce has been seen before within the skew window
+7. Insert the nonce into the replay cache on success
+8. Respond with `OK\n` on success or `ERR\n` on failure
+
+### Nonce replay detection
+
+The SSH server maintains a per-process `NonceCache` (`HashMap<String, Instant>` behind `Arc<Mutex<...>>`) that tracks nonces seen within the handshake skew window. A background tokio task reaps expired entries every 60 seconds. If a valid preface is presented with a previously-seen nonce, the handshake is rejected. This prevents replay attacks within the timestamp validity window.
 
 ### HMAC computation
 
@@ -517,7 +523,12 @@ This function is shared between the CLI and TUI via the `navigator-core::forward
 
 1. **mTLS (transport layer)** -- when TLS is configured, the CLI authenticates to the gateway using client certificates. The `ssh-proxy` subprocess inherits TLS options from the parent CLI process.
 2. **Session token (application layer)** -- the gateway validates the session token against the persistence layer. Tokens are scoped to a specific sandbox and can be revoked.
-3. **NSSH1 handshake (gateway-to-sandbox)** -- the shared handshake secret proves the connection originated from an authorized gateway. The timestamp + nonce prevent replay attacks within the skew window.
+3. **NSSH1 handshake (gateway-to-sandbox)** -- the shared handshake secret proves the connection originated from an authorized gateway. The timestamp + nonce prevent replay attacks within the skew window. The nonce replay cache rejects duplicates.
+4. **Kubernetes NetworkPolicy** -- a Helm-managed `NetworkPolicy` restricts ingress to sandbox pods on port 2222 to only the gateway pod, preventing lateral movement from other in-cluster workloads. Controlled by `networkPolicy.enabled` in the Helm values (default: `true`).
+
+### Mandatory handshake secret
+
+The NSSH1 handshake secret (`NEMOCLAW_SSH_HANDSHAKE_SECRET`) is required. Both the server and sandbox will refuse to start if the secret is empty or unset. For cluster deployments the secret is auto-generated by the entrypoint script (`deploy/docker/cluster-entrypoint.sh`) via `openssl rand -hex 32` and injected into the Helm values.
 
 ### What SSH auth does NOT enforce
 
@@ -542,7 +553,7 @@ The sandbox generates a fresh Ed25519 host key on every startup. The CLI disable
 | `ssh_gateway_port`         | `8080`           | Public port for gateway connections (0 = use bind port) |
 | `ssh_connect_path`         | `/connect/ssh`   | HTTP path for CONNECT requests |
 | `sandbox_ssh_port`         | `2222`           | SSH listen port inside sandbox pods |
-| `ssh_handshake_secret`     | (empty)          | Shared HMAC key for NSSH1 handshake |
+| `ssh_handshake_secret`     | (required)       | Shared HMAC key for NSSH1 handshake (server fails to start if empty) |
 | `ssh_handshake_skew_secs`  | `300`            | Maximum allowed clock skew (seconds) |
 
 ### Sandbox environment variables

crates/navigator-sandbox/src/lib.rs

@@ -404,7 +404,14 @@ pub async fn run_sandbox(
         let addr: SocketAddr = listen_addr.parse().into_diagnostic()?;
         let policy_clone = policy.clone();
         let workdir_clone = workdir.clone();
-        let secret = ssh_handshake_secret.unwrap_or_default();
+        let secret = ssh_handshake_secret
+            .filter(|s| !s.is_empty())
+            .ok_or_else(|| {
+                miette::miette!(
+                    "NEMOCLAW_SSH_HANDSHAKE_SECRET is required when SSH is enabled.\n\
+                     Set --ssh-handshake-secret or the NEMOCLAW_SSH_HANDSHAKE_SECRET env var."
+                )
+            })?;
         let proxy_url = ssh_proxy_url;
         let netns_fd = ssh_netns_fd;
         let ca_paths = ca_file_paths.clone();

crates/navigator-sandbox/src/ssh.rs

@@ -21,14 +21,19 @@ use std::net::SocketAddr;
 use std::os::fd::{AsRawFd, RawFd};
 use std::path::PathBuf;
 use std::process::Command;
-use std::sync::{Arc, mpsc};
-use std::time::{Duration, SystemTime, UNIX_EPOCH};
+use std::sync::{Arc, Mutex, mpsc};
+use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
 use tokio::io::{AsyncReadExt, AsyncWriteExt};
 use tokio::net::TcpListener;
 use tracing::{info, warn};
 
 const PREFACE_MAGIC: &str = "NSSH1";
 
+/// A time-bounded set of nonces used to detect replayed NSSH1 handshakes.
+/// Each entry records the `Instant` it was inserted; a background reaper task
+/// periodically evicts entries older than the handshake skew window.
+type NonceCache = Arc<Mutex<HashMap<String, Instant>>>;
+
 /// Perform SSH server initialization: generate a host key, build the config,
 /// and bind the TCP listener. Extracted so that startup errors can be forwarded
 /// through the readiness channel rather than being silently logged.
@@ -85,6 +90,23 @@ pub async fn run_ssh_server(
         }
     };
 
+    // Nonce cache for replay detection. Entries are evicted by a background
+    // reaper once they exceed the handshake skew window.
+    let nonce_cache: NonceCache = Arc::new(Mutex::new(HashMap::new()));
+
+    // Background task that periodically purges expired nonces.
+    let reaper_cache = nonce_cache.clone();
+    let ttl = Duration::from_secs(handshake_skew_secs);
+    tokio::spawn(async move {
+        let mut interval = tokio::time::interval(Duration::from_secs(60));
+        loop {
+            interval.tick().await;
+            if let Ok(mut cache) = reaper_cache.lock() {
+                cache.retain(|_, inserted| inserted.elapsed() < ttl);
+            }
+        }
+    });
+
     loop {
         let (stream, peer) = listener.accept().await.into_diagnostic()?;
         stream.set_nodelay(true).into_diagnostic()?;
@@ -95,6 +117,7 @@ pub async fn run_ssh_server(
         let proxy_url = proxy_url.clone();
         let ca_paths = ca_paths.clone();
         let provider_env = provider_env.clone();
+        let nonce_cache = nonce_cache.clone();
 
         tokio::spawn(async move {
             if let Err(err) = handle_connection(
@@ -109,6 +132,7 @@ pub async fn run_ssh_server(
                 proxy_url,
                 ca_paths,
                 provider_env,
+                &nonce_cache,
             )
             .await
             {
@@ -131,12 +155,13 @@ async fn handle_connection(
     proxy_url: Option<String>,
     ca_file_paths: Option<Arc<(PathBuf, PathBuf)>>,
     provider_env: HashMap<String, String>,
+    nonce_cache: &NonceCache,
 ) -> Result<()> {
     info!(peer = %peer, "SSH connection: reading handshake preface");
     let mut line = String::new();
     read_line(&mut stream, &mut line).await?;
     info!(peer = %peer, preface_len = line.len(), "SSH connection: preface received, verifying");
-    if !verify_preface(&line, secret, handshake_skew_secs)? {
+    if !verify_preface(&line, secret, handshake_skew_secs, nonce_cache)? {
         warn!(peer = %peer, "SSH connection: handshake verification failed");
         let _ = stream.write_all(b"ERR\n").await;
         return Ok(());
@@ -178,7 +203,12 @@ async fn read_line(stream: &mut tokio::net::TcpStream, buf: &mut String) -> Resu
     Ok(())
 }
 
-fn verify_preface(line: &str, secret: &str, handshake_skew_secs: u64) -> Result<bool> {
+fn verify_preface(
+    line: &str,
+    secret: &str,
+    handshake_skew_secs: u64,
+    nonce_cache: &NonceCache,
+) -> Result<bool> {
     let parts: Vec<&str> = line.split_whitespace().collect();
     if parts.len() != 5 || parts[0] != PREFACE_MAGIC {
         return Ok(false);
@@ -202,7 +232,22 @@ fn verify_preface(line: &str, secret: &str, handshake_skew_secs: u64) -> Result<
 
     let payload = format!("{token}|{timestamp}|{nonce}");
     let expected = hmac_sha256(secret.as_bytes(), payload.as_bytes());
-    Ok(signature == expected)
+    if signature != expected {
+        return Ok(false);
+    }
+
+    // Reject replayed nonces. The cache is bounded by the reaper task which
+    // evicts entries older than `handshake_skew_secs`.
+    let mut cache = nonce_cache
+        .lock()
+        .map_err(|_| miette::miette!("nonce cache lock poisoned"))?;
+    if cache.contains_key(nonce) {
+        warn!(nonce = nonce, "NSSH1 nonce replay detected");
+        return Ok(false);
+    }
+    cache.insert(nonce.to_string(), Instant::now());
+
+    Ok(true)
 }
 
 fn hmac_sha256(key: &[u8], data: &[u8]) -> String {
@@ -1085,4 +1130,103 @@ mod tests {
             "expected all 100 KiB delivered before EOF"
         );
     }
+
+    // -----------------------------------------------------------------------
+    // verify_preface tests
+    // -----------------------------------------------------------------------
+
+    /// Build a valid NSSH1 preface line with the given parameters.
+    fn build_preface(token: &str, secret: &str, nonce: &str, timestamp: i64) -> String {
+        let payload = format!("{token}|{timestamp}|{nonce}");
+        let signature = hmac_sha256(secret.as_bytes(), payload.as_bytes());
+        format!("{PREFACE_MAGIC} {token} {timestamp} {nonce} {signature}")
+    }
+
+    fn fresh_nonce_cache() -> NonceCache {
+        Arc::new(Mutex::new(HashMap::new()))
+    }
+
+    fn current_timestamp() -> i64 {
+        i64::try_from(
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .unwrap()
+                .as_secs(),
+        )
+        .unwrap()
+    }
+
+    #[test]
+    fn verify_preface_accepts_valid_preface() {
+        let secret = "test-secret-key";
+        let nonce = "unique-nonce-1";
+        let ts = current_timestamp();
+        let line = build_preface("tok1", secret, nonce, ts);
+        let cache = fresh_nonce_cache();
+
+        assert!(verify_preface(&line, secret, 300, &cache).unwrap());
+    }
+
+    #[test]
+    fn verify_preface_rejects_replayed_nonce() {
+        let secret = "test-secret-key";
+        let nonce = "replay-nonce";
+        let ts = current_timestamp();
+        let line = build_preface("tok1", secret, nonce, ts);
+        let cache = fresh_nonce_cache();
+
+        // First attempt should succeed.
+        assert!(verify_preface(&line, secret, 300, &cache).unwrap());
+        // Second attempt with the same nonce should be rejected.
+        assert!(!verify_preface(&line, secret, 300, &cache).unwrap());
+    }
+
+    #[test]
+    fn verify_preface_rejects_expired_timestamp() {
+        let secret = "test-secret-key";
+        let nonce = "expired-nonce";
+        // Timestamp 600 seconds in the past, with a 300-second skew window.
+        let ts = current_timestamp() - 600;
+        let line = build_preface("tok1", secret, nonce, ts);
+        let cache = fresh_nonce_cache();
+
+        assert!(!verify_preface(&line, secret, 300, &cache).unwrap());
+    }
+
+    #[test]
+    fn verify_preface_rejects_invalid_hmac() {
+        let secret = "test-secret-key";
+        let nonce = "hmac-nonce";
+        let ts = current_timestamp();
+        // Build with the correct secret, then verify with the wrong one.
+        let line = build_preface("tok1", secret, nonce, ts);
+        let cache = fresh_nonce_cache();
+
+        assert!(!verify_preface(&line, "wrong-secret", 300, &cache).unwrap());
+    }
+
+    #[test]
+    fn verify_preface_rejects_malformed_input() {
+        let cache = fresh_nonce_cache();
+
+        // Too few parts.
+        assert!(!verify_preface("NSSH1 tok1 123", "s", 300, &cache).unwrap());
+        // Wrong magic.
+        assert!(!verify_preface("NSSH2 tok1 123 nonce sig", "s", 300, &cache).unwrap());
+        // Empty string.
+        assert!(!verify_preface("", "s", 300, &cache).unwrap());
+    }
+
+    #[test]
+    fn verify_preface_distinct_nonces_both_accepted() {
+        let secret = "test-secret-key";
+        let ts = current_timestamp();
+        let cache = fresh_nonce_cache();
+
+        let line1 = build_preface("tok1", secret, "nonce-a", ts);
+        let line2 = build_preface("tok1", secret, "nonce-b", ts);
+
+        assert!(verify_preface(&line1, secret, 300, &cache).unwrap());
+        assert!(verify_preface(&line2, secret, 300, &cache).unwrap());
+    }
 }

crates/navigator-server/src/lib.rs

@@ -92,6 +92,11 @@ pub async fn run_server(config: Config, tracing_log_bus: TracingLogBus) -> Resul
     if database_url.is_empty() {
         return Err(Error::config("database_url is required"));
     }
+    if config.ssh_handshake_secret.is_empty() {
+        return Err(Error::config(
+            "ssh_handshake_secret is required. Set --ssh-handshake-secret or NEMOCLAW_SSH_HANDSHAKE_SECRET",
+        ));
+    }
 
     let store = Store::connect(database_url).await?;
     let sandbox_client = SandboxClient::new(

crates/navigator-server/src/sandbox/mod.rs

@@ -915,9 +915,7 @@ fn apply_required_env(
     if !ssh_listen_addr.is_empty() {
         upsert_env(env, "NEMOCLAW_SSH_LISTEN_ADDR", ssh_listen_addr);
     }
-    if !ssh_handshake_secret.is_empty() {
-        upsert_env(env, "NEMOCLAW_SSH_HANDSHAKE_SECRET", ssh_handshake_secret);
-    }
+    upsert_env(env, "NEMOCLAW_SSH_HANDSHAKE_SECRET", ssh_handshake_secret);
     upsert_env(
         env,
         "NEMOCLAW_SSH_HANDSHAKE_SKEW_SECS",
@@ -1219,4 +1217,29 @@ mod tests {
 
         assert_eq!(derive_phase(&status, false), SandboxPhase::Ready);
     }
+
+    #[test]
+    fn apply_required_env_always_injects_ssh_handshake_secret() {
+        let mut env = Vec::new();
+        apply_required_env(
+            &mut env,
+            "sandbox-1",
+            "my-sandbox",
+            "https://endpoint:8080",
+            "0.0.0.0:2222",
+            "my-secret-value",
+            300,
+        );
+
+        let secret_entry = env
+            .iter()
+            .find(|e| {
+                e.get("name").and_then(|v| v.as_str()) == Some("NEMOCLAW_SSH_HANDSHAKE_SECRET")
+            })
+            .expect("NEMOCLAW_SSH_HANDSHAKE_SECRET must be present in env");
+        assert_eq!(
+            secret_entry.get("value").and_then(|v| v.as_str()),
+            Some("my-secret-value")
+        );
+    }
 }

deploy/docker/cluster-entrypoint.sh

@@ -249,6 +249,11 @@ if [ -n "${IMAGE_PULL_POLICY:-}" ] && [ -f "$HELMCHART" ]; then
     sed -i "s|pullPolicy: Always|pullPolicy: ${IMAGE_PULL_POLICY}|" "$HELMCHART"
 fi
 
+# Generate a random SSH handshake secret for the NSSH1 HMAC handshake between
+# the gateway and sandbox SSH servers. This is required — the server will refuse
+# to start without it.
+SSH_HANDSHAKE_SECRET="${SSH_HANDSHAKE_SECRET:-$(openssl rand -hex 32)}"
+
 # Inject SSH gateway host/port into the HelmChart manifest so the navigator
 # server returns the correct address to CLI clients for SSH proxy CONNECT.
 if [ -f "$HELMCHART" ]; then
@@ -266,6 +271,8 @@ if [ -f "$HELMCHART" ]; then
         # Clear the placeholder so the default (8080) is used
         sed -i "s|sshGatewayPort: __SSH_GATEWAY_PORT__|sshGatewayPort: 0|g" "$HELMCHART"
     fi
+    echo "Setting SSH handshake secret"
+    sed -i "s|__SSH_HANDSHAKE_SECRET__|${SSH_HANDSHAKE_SECRET}|g" "$HELMCHART"
 fi
 
 # Inject chart checksum into the HelmChart manifest so that a changed chart

deploy/helm/navigator/templates/networkpolicy.yaml

@@ -0,0 +1,32 @@
+{{- if .Values.networkPolicy.enabled }}
+# NetworkPolicy restricting SSH ingress on sandbox pods to the gateway pod.
+# Sandbox pods are dynamically created by the server and labelled with
+# navigator.ai/managed-by=navigator. This policy ensures only the gateway
+# (navigator server) pod can reach the sandbox SSH port (2222), blocking
+# lateral movement from other in-cluster workloads.
+apiVersion: networking.k8s.io/v1
+kind: NetworkPolicy
+metadata:
+  name: {{ include "navigator.fullname" . }}-sandbox-ssh
+  namespace: {{ .Values.server.sandboxNamespace }}
+  labels:
+    {{- include "navigator.labels" . | nindent 4 }}
+spec:
+  podSelector:
+    matchLabels:
+      navigator.ai/managed-by: navigator
+  policyTypes:
+    - Ingress
+  ingress:
+    - from:
+        - namespaceSelector:
+            matchLabels:
+              kubernetes.io/metadata.name: {{ .Release.Namespace }}
+          podSelector:
+            matchLabels:
+              app.kubernetes.io/name: {{ include "navigator.name" . }}
+              app.kubernetes.io/instance: {{ .Release.Name }}
+      ports:
+        - protocol: TCP
+          port: 2222
+{{- end }}

deploy/helm/navigator/templates/statefulset.yaml

@@ -58,6 +58,8 @@ spec:
             - name: NEMOCLAW_SSH_GATEWAY_PORT
               value: {{ .Values.server.sshGatewayPort | quote }}
             {{- end }}
+            - name: NEMOCLAW_SSH_HANDSHAKE_SECRET
+              value: {{ required "server.sshHandshakeSecret is required" .Values.server.sshHandshakeSecret | quote }}
             - name: NEMOCLAW_TLS_CERT
               value: /etc/navigator-tls/server/tls.crt
             - name: NEMOCLAW_TLS_KEY

deploy/helm/navigator/values.yaml

@@ -77,6 +77,10 @@ server:
   sshGatewayPort: 0
   # TLS configuration for the server. The server always terminates mTLS
   # directly and requires client certificates.
+  # HMAC secret used for the NSSH1 handshake between gateway and sandbox SSH.
+  # Required — the server will refuse to start if empty. For cluster deployments
+  # this is auto-generated by the entrypoint script.
+  sshHandshakeSecret: ""
   tls:
     # K8s secret (type kubernetes.io/tls) with tls.crt and tls.key for the server
     certSecretName: navigator-server-tls
@@ -84,3 +88,7 @@ server:
     clientCaSecretName: navigator-server-client-ca
     # K8s secret mounted into sandbox pods for mTLS to the server
     clientTlsSecretName: navigator-client-tls
+
+# NetworkPolicy restricting SSH ingress on sandbox pods to the gateway only.
+networkPolicy:
+  enabled: true

deploy/kube/manifests/navigator-helmchart.yaml

@@ -31,6 +31,7 @@ spec:
       sandboxImage: d1i0nduu2f6qxk.cloudfront.net/navigator/sandbox:latest
       sshGatewayHost: __SSH_GATEWAY_HOST__
       sshGatewayPort: __SSH_GATEWAY_PORT__
+      sshHandshakeSecret: __SSH_HANDSHAKE_SECRET__
       grpcEndpoint: "https://navigator.navigator.svc.cluster.local:8080"
       tls:
         certSecretName: navigator-server-tls

tasks/scripts/cluster-deploy-fast.sh

@@ -379,6 +379,13 @@ if [[ "${needs_helm_upgrade}" == "1" ]]; then
   # terminates mTLS (there is no server.tls.enabled toggle). Without this,
   # a prior Helm override or chart default change could silently regress
   # sandbox callbacks to plaintext.
+  # Retrieve the existing handshake secret from the running release, or generate
+  # a new one if this is the first deploy with the mandatory secret.
+  EXISTING_SECRET=$(helm get values navigator -n navigator -o json 2>/dev/null \
+    | grep -o '"sshHandshakeSecret":"[^"]*"' \
+    | cut -d'"' -f4) || true
+  SSH_HANDSHAKE_SECRET="${EXISTING_SECRET:-$(openssl rand -hex 32)}"
+
   helm upgrade navigator deploy/helm/navigator \
     --namespace navigator \
     --set image.repository=${IMAGE_REPO_BASE}/server \
@@ -389,6 +396,7 @@ if [[ "${needs_helm_upgrade}" == "1" ]]; then
     --set server.tls.certSecretName=navigator-server-tls \
     --set server.tls.clientCaSecretName=navigator-server-client-ca \
     --set server.tls.clientTlsSecretName=navigator-client-tls \
+    --set server.sshHandshakeSecret=${SSH_HANDSHAKE_SECRET} \
     "${helm_wait_args[@]}"
   helm_end=$(date +%s)
   log_duration "Helm upgrade" "${helm_start}" "${helm_end}"